
Read-only Bash audit script that triages CVE-2026-99999 (PwnEscape) Linux kernel IPv4 checksum OOB write, checking patch status, module reachability, and risk score.
Read-only audit script for CVE-2026-99999 (a.k.a. PwnEscape). The bug is in the Linux kernel's IPv4 stack: it trusts the checksum metadata a packet arrives with. An untrusted VM guest can use it to escape to the host, and a local user or container can use it for privilege escalation.
___ ___
| _ \__ __ ___ _ | __|___ __ __ _ _ __ ___
| _/\ V V / ' \ | _|(_-</ _/ _` | '_ \/ -_)
|_| \_/\_/|_||_||___/__/\__\__,_| .__/\___|
|_|
CVE-2026-99999 ipv4 skb_checksum_help() csum_start OOB write
A CHECKSUM_PARTIAL packet carries csum_start/csum_offset taken from a virtio_net_hdr. The packet source controls those values. When the IPv4 stack later completes the checksum in skb_checksum_help(), nothing has checked the offsets against the IPv4 header it has already validated. The checksum write then lands on the network header or past it (CWE-787). Red Hat lists the outcomes as code execution on the host, a crash, or a leak of kernel memory.
The upstream fix has two Fixes: tags, one for each way bad metadata gets in. The script checks both:
| Entry | Introduced by | Who can reach it |
|---|---|---|
tun/tap (also macvtap, ipvtap) with virtio_net_hdr | f43798c27684 "tun: Allow GSO using virtio_net_hdr" (2.6.27) | Guests of QEMU/KVM, vhost-net, Firecracker, Cloud Hypervisor; any user who owns a persistent tap; CAP_NET_ADMIN via userns |
AF_PACKET with PACKET_VNET_HDR | bfd5f4a3d605 "packet: Add GSO/csum offload support." | Anyone with CAP_NET_RAW: unprivileged userns, Docker/Podman containers by default |
It is the host kernel that has to be patched. Updating guest kernels does nothing.
| CVE | CVE-2026-99999 |
| Component | Linux kernel net/ipv4 / skb_checksum_help() |
| Class | Out-of-bounds write → VM escape / LPE |
| Reporter | Paulos Yibelo (pwn.ai), Vercel Sandbox challenge |
| Upstream fix | 6785011f8b16 "ipv4: validate checksum_start before completing checksum", in 7.3-rc7 |
| Stable fixes | 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.57, 7.2.11 |
| Red Hat | Important, CVSS 8.8 (AV:L/PR:L/S:C); no runtime mitigation. Fixed: RHEL 6 ELS 2.6.32-754.66.1.el6, RHEL 7 ELS 3.10.0-1160.167.2.el7, RHEL 8 4.18.0-553.178.1.el8_10, RHEL 9 5.14.0-687.59.1.el9_8, RHEL 10 6.12.0-211.67.1.el10_2; EUS/AUS streams carry other builds |
| Ubuntu | CVSS 9.3 (AV:L/PR:N/S:C); source package linux: 26.10 7.3.0-10.10, 26.04 7.0.0-43.43, 24.04 6.8.0-149.149, 22.04 5.15.0-203.213; 20.04/18.04/16.04 pending, 14.04 gets no fix; flavours and HWE kernels are fixed in their own builds |
| Debian | bookworm fixed in 6.1.190-1 (DLA-4841-1), trixie in 6.12.113-1 (DSA-6568-1), unstable in 7.2.11-1 |
The stable versions were checked against the kernel.org ChangeLog-* files on 2026-10-17. In each branch, the listed release is the first to carry the fix and the release before it lacks it. Distro states were re-read on 2026-11-16 (Debian security tracker, ubuntu.com/security, Red Hat security data API). The report prints the age of this snapshot and warns once it passes 14 days (matrix_date, matrix_age_days, matrix_stale in JSON; a second NOTE: line with --quiet).
The upstream commit calls itself defence in depth: "a more complete validation at input is forthcoming". A follow-up virtio-net hardening series may get its own CVE.
It only triages and changes nothing. It never loads a module, never touches sysctl, and never runs a PoC, so it is safe to run in production.
changelog.Debian.gz or rpm -q --changelog) for the CVE ID or the upstream commit subject. A fix found there always wins. On Ubuntu the kernel changelog is read from linux-modules-$(uname -r); the signed linux-image-* package only carries a signing wrapper. Ubuntu keeps the last 10 uploads there, so for the linux source package a version threshold decides as well.tun, macvtap, ipvtap, af_packet) for being:
lsmod)/boot/config-$(uname -r), /proc/config.gz, modules.builtin), or compiled outmodules.dep; trusted only when the module tree exists, so a container isn't mistaken for a safe host)install <mod> /bin/true|false override, by a blacklist that the kernel's autoload alias honours, or by kernel.modules_disabled=1/dev/kvm, hypervisor and vhost processes, tun/tap interfaces, user-owned persistent tap devices (no capability needed), container runtimes.skb_checksum_help / bad-offload warnings (--since / --until).PATCHED / NOT_APPLICABLE / MITIGATED / VULNERABLE / UNKNOWN.PATCHED: the changelog lists the fix, the package is at or above a concrete vendor threshold, or the kernel is an upstream-tracking one at or above the stable/mainline fix.
VULNERABLE: the kernel is unfixed and at least one entry path is reachable.
MITIGATED: the kernel is unfixed, but both paths are closed:
tun, macvtap, ipvtap) is loaded, built in, or loadable.A module that is already loaded still counts, because a modprobe rule doesn't unload it.
UNKNOWN: the vendor has no fixed build yet and the changelog could not be read. Also used inside a container, where the package data describes the image, not the host kernel.
The AppArmor userns restriction lowers the score but is not counted as a block.
sudo ./cve-2026-99999-audit.sh # full text report
sudo ./cve-2026-99999-audit.sh --json # machine-readable JSON, no banner
sudo ./cve-2026-99999-audit.sh --quiet --no-hunt # one-line verdict for mass scans
sudo ./cve-2026-99999-audit.sh --since 2026-09-01 # widen IoC hunt window
sudo ./cve-2026-99999-audit.sh --since 2026-09-22 --until 2026-10-17 # bounded IoC window
sudo ./cve-2026-99999-audit.sh --no-banner # text report without the ASCII banner
./cve-2026-99999-audit.sh --version # print script version and exit
The script also runs without root. It then can't read every changelog or sysfs entry, and gives UNKNOWN more often.